feat(gpt-bot): Replace with SmolLM2 family and fix model switching

- Replace Qwen/Llama models with SmolLM2 family (135M, 360M, 1.7B)
- Fix model switching bug by calling dispose() before loading new model
- Add RAM detection (navigator.deviceMemory) to auto-select best model
- Dynamically populate model selector dropdown
- Update architecture specs and documentation for SmolLM2

Author: jeremymanning

demos/chatbot-evolution/README.md

@@ -120,28 +120,28 @@ An interactive journey through 60 years of conversational AI development, from E
 - Follow-up questions to test context
 - Compare with rule-based responses above
 
-### 5. GPT & Transformers (2020s)
-**Innovation:** Large Language Models with transformers
+### 5. SmolLM2 & Transformers (2020s)
+**Creator:** HuggingFace (2024)
 
-**Approach:** Self-attention, massive scale, pre-training
+**Approach:** Decoder-only transformer, optimized for browser/edge deployment
 
 **How it works:**
-- Transformer architecture (self-attention)
-- Pre-trained on vast text corpora
-- Fine-tuned for instruction following
-- Few-shot learning capabilities
+- Auto-selects model size based on your device RAM
+- Available in 135M, 360M, and 1.7B parameter variants
+- Pre-trained on FineWeb-Edu, code, and synthetic data
+- Instruction-tuned for helpful conversations
+- Runs entirely in your browser via WebGPU or WASM
 
 **Famous for:**
-- ChatGPT, GPT-4, Claude, Gemini
-- Unprecedented fluency and coherence
-- Reasoning and problem-solving
-- Multi-turn conversations
+- Best-in-class quality for browser-sized models
+- Native Transformers.js support (ONNX bundled)
+- Open-source (Apache 2.0)
 
 **Capabilities:**
-- Context understanding
-- Knowledge integration
-- Creative generation
-- Task completion
+- Multi-turn conversation with context preservation
+- General knowledge and reasoning
+- Code generation and explanation
+- Helpful, instruction-following responses
 
 ## Features
 
@@ -237,7 +237,7 @@ An interactive journey through 60 years of conversational AI development, from E
 - Initial model loading time in browser
 
 **What Makes This Demo Special:**
-This demo runs **real neural models** (BlenderBot 90M, Qwen2.5 0.5B) directly in your browser using Transformers.js. You'll experience genuine neural behavior - not simulations - allowing you to see the clear evolution from rule-based to learned conversation.
+This demo runs **real neural models** (BlenderBot 90M, SmolLM2 135M-1.7B) directly in your browser using Transformers.js. The SmolLM2 model auto-selects based on your device RAM. You'll experience genuine neural behavior - not simulations - allowing you to see the clear evolution from rule-based to learned conversation.
 
 ## Historical Timeline
 
@@ -263,7 +263,7 @@ This demo runs **real neural models** (BlenderBot 90M, Qwen2.5 0.5B) directly in
 - **PARRY**: State machine with emotional variables
 - **A.L.I.C.E.**: ~40,000 AIML patterns
 - **BlenderBot Small**: 90 million parameters (real neural model)
-- **Qwen2.5 0.5B**: 500 million parameters (real neural model)
+- **SmolLM2**: 135M-1.7B parameters (real neural model, auto-selected)
 - **GPT-3**: 175 billion parameters (comparison reference)
 
 ### Architectural Progression
@@ -340,11 +340,13 @@ AIML-inspired pattern matching with improved context handling over ELIZA.
 - Loads and runs entirely in your browser (may take 30-60 seconds initially)
 - Fallback to DialoGPT-small if BlenderBot fails to load
 
-### GPT / Qwen2.5
-Uses Transformers.js with Qwen2.5-0.5B-Instruct (500M parameters) for actual neural text generation in the browser. Falls back to Llama-3.2-1B-Instruct if Qwen fails to load. Demonstrates modern transformer capabilities:
+### SmolLM2
+Uses Transformers.js with HuggingFace's SmolLM2 family (135M, 360M, 1.7B parameters) for actual neural text generation in the browser. Key features:
+- **Auto-selects model** based on device RAM (navigator.deviceMemory API)
+- **Proper model switching** with dispose() to prevent memory leaks
 - Instruction-tuned for natural conversations
 - Conversation history preserved across turns
-- Runs entirely in-browser via WASM or WebGPU
+- Runs entirely in-browser via WebGPU (preferred) or WASM fallback
 - May take 30-60 seconds to download on first load
 
 ## Extensions

demos/chatbot-evolution/index.html

@@ -41,7 +41,7 @@ <h1 class="hero-title">Chatbot Evolution Timeline</h1>
                     <span class="era-label">2020<br>BlenderBot</span>
                 </div>
                 <div class="era era-2020s" data-era="2020s">
-                    <span class="era-label">2024<br>Qwen2.5</span>
+                    <span class="era-label">2024<br>SmolLM2</span>
                 </div>
             </div>
         </section>
@@ -455,20 +455,20 @@ <h2>2020s: GPT & Transformers</h2>
 
                     <div class="chatbot-info">
                         <div class="info-card">
-                            <h3>About Modern LLMs</h3>
-                            <p><strong>Innovation:</strong> Instruction-tuned transformers</p>
+                            <h3>About SmolLM2</h3>
+                            <p><strong>Creator:</strong> HuggingFace (2024)</p>
                             <p><strong>Method:</strong> Decoder-only transformer with chat templates</p>
-                            <p><strong>Models:</strong> Qwen2.5 0.5B, SmolLM 360M</p>
-                            <p><strong>Context:</strong> Multi-turn conversation support</p>
+                            <p><strong>Models:</strong> 135M, 360M, 1.7B parameters</p>
+                            <p><strong>Innovation:</strong> Optimized for browser/edge deployment</p>
                         </div>
 
                         <div class="info-card">
                             <h3>How It Works</h3>
                             <ul>
-                                <li>Pre-trained on vast text corpora</li>
-                                <li>Fine-tuned for instruction following</li>
-                                <li>System prompts guide behavior</li>
-                                <li>Runs entirely in your browser (WASM/WebGPU)</li>
+                                <li>Auto-selects model based on your device RAM</li>
+                                <li>Pre-trained on web text, code, and reasoning data</li>
+                                <li>Instruction-tuned for helpful conversations</li>
+                                <li>Runs entirely in your browser (WebGPU/WASM)</li>
                             </ul>
                         </div>
                     </div>
@@ -490,24 +490,21 @@ <h3>How It Works</h3>
                             <div class="chat-messages" id="gpt-messages"></div>
                             <div class="chat-input-area">
                                 <div class="model-selector-row">
-                                    <select id="gpt-model-selector" class="model-selector">
-                                        <option value="0">Qwen 2.5 0.5B (Alibaba)</option>
-                                        <option value="1">Llama 3.2 1B (Meta)</option>
-                                    </select>
+                                    <select id="gpt-model-selector" class="model-selector"></select>
                                     <button class="btn-clear-chat" id="gpt-clear-btn" title="Clear chat history">Clear</button>
                                 </div>
                                 <input type="text" class="chat-input" id="gpt-input" placeholder="Talk to the model...">
                                 <button class="chat-send" id="gpt-send-btn" onclick="sendMessage('gpt')">Send</button>
                             </div>
-                            <p class="demo-note">Select a model above. Conversation history is preserved. First message loads the model (~30s).</p>
+                            <p class="demo-note">Model auto-selected based on your device RAM. First message loads the model (~30-60s).</p>
                         </div>
                     </div>
 
                     <!-- Architecture Tab -->
                     <div class="chatbot-tab-content" id="gpt-architecture-tab">
                         <div class="architecture-content">
                             <div class="architecture-diagram">
-                                <h4>Decoder-Only Transformer (Qwen2.5 / Llama 3.2)</h4>
+                                <h4>Decoder-Only Transformer (SmolLM2)</h4>
                                 <div class="arch-flow">
                                     <div class="arch-block input-block">
                                         <div class="block-label">Input Prompt</div>
@@ -523,9 +520,9 @@ <h4>Decoder-Only Transformer (Qwen2.5 / Llama 3.2)</h4>
                                         <div class="block-label">Transformer Decoder Stack</div>
                                         <div class="block-content">
                                             <div class="sub-block">Grouped-Query Attention</div>
-                                            <div class="sub-block">SiLU FFN</div>
+                                            <div class="sub-block">SwiGLU FFN</div>
                                             <div class="sub-block">RMSNorm</div>
-                                            <div class="block-note">x24 layers (Qwen) / x16 layers (Llama)</div>
+                                            <div class="block-note">x9 (135M) / x16 (360M) / x24 (1.7B) layers</div>
                                         </div>
                                     </div>
                                     <div class="arch-arrow">&#8595;</div>
@@ -559,21 +556,21 @@ <h5>Quantization (q4)</h5>
                             </div>
 
                             <div class="model-specs">
-                                <h4>Available Models</h4>
+                                <h4>SmolLM2 Model Family</h4>
                                 <table class="specs-table">
-                                    <tr><th>Spec</th><th>Qwen 2.5 0.5B</th><th>Llama 3.2 1B</th></tr>
-                                    <tr><td>Parameters</td><td>500 Million</td><td>1 Billion</td></tr>
-                                    <tr><td>Architecture</td><td>Decoder-Only</td><td>Decoder-Only</td></tr>
-                                    <tr><td>Layers</td><td>24</td><td>16</td></tr>
-                                    <tr><td>Hidden Size</td><td>896</td><td>2048</td></tr>
-                                    <tr><td>Context Length</td><td>32,768 tokens</td><td>131,072 tokens</td></tr>
-                                    <tr><td>Organization</td><td>Alibaba</td><td>Meta</td></tr>
-                                    <tr><td>Year</td><td>2024</td><td>2024</td></tr>
+                                    <tr><th>Spec</th><th>135M</th><th>360M</th><th>1.7B</th></tr>
+                                    <tr><td>Parameters</td><td>135 Million</td><td>360 Million</td><td>1.7 Billion</td></tr>
+                                    <tr><td>Layers</td><td>9</td><td>16</td><td>24</td></tr>
+                                    <tr><td>Hidden Size</td><td>576</td><td>960</td><td>2048</td></tr>
+                                    <tr><td>Download (q4)</td><td>~85 MB</td><td>~210 MB</td><td>~980 MB</td></tr>
+                                    <tr><td>Min RAM</td><td>2 GB</td><td>4 GB</td><td>8 GB</td></tr>
+                                    <tr><td>Context Length</td><td colspan="3">8,192 tokens</td></tr>
                                 </table>
                                 <div class="model-note">
                                     <strong>Model Links:</strong> 
-                                    <a href="https://huggingface.co/onnx-community/Qwen2.5-0.5B-Instruct" target="_blank">Qwen2.5-0.5B-Instruct</a> | 
-                                    <a href="https://huggingface.co/onnx-community/Llama-3.2-1B-Instruct" target="_blank">Llama-3.2-1B-Instruct</a>
+                                    <a href="https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct" target="_blank">135M</a> | 
+                                    <a href="https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct" target="_blank">360M</a> | 
+                                    <a href="https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct" target="_blank">1.7B</a>
                                 </div>
                             </div>
                         </div>
@@ -642,8 +639,8 @@ <h3>Evolution Stats</h3>
                             <span class="stat-value">90M params</span>
                         </div>
                         <div class="stat-item">
-                            <span class="stat-label">Qwen 2.5 (2024)</span>
-                            <span class="stat-value">500M params</span>
+                            <span class="stat-label">SmolLM2 (2024)</span>
+                            <span class="stat-value">135M-1.7B params</span>
                         </div>
                         <div class="stat-item">
                             <span class="stat-label">Claude 4 Opus (2025)</span>